Walking Around Your Nearest Neighbors

on YouTube ↗06m

" summary

Builds a streaming nearest-neighbor pipeline that keeps item embeddings out of the search index by computing neighbors on ingest, updating neighbors of affected items, and caching results for fast recommendations. Uses a lightweight app with Feist and a daily batch to correct model and embedding-space drift, following a lambda architecture with speed and batch layers. Claims a production deployment that can brute-force rank up to 5 million records in under 10 ms, enabling immediate recommendations for newly posted items.

AI-written summary · about

" chapters (7)

1. 0:00Start
2. 0:15Walking around your nearest neighbors
3. 0:39Use case: Recommendations @ OLX Group
4. 1:19Constraints
5. 1:43Embedding space
6. 3:00Flow
7. 4:36Lambda architecture

" related talks

• 38mEmbeddings! Embeddings everywhere!Maciej Arciuch, Karol Grzegorczyk · PyData London 2019
• 38mSearch: A New EraDavid Pilato · Big Data Conference Europe 2023
• 20mDataEngBytes 2023 - PER-T1-06 - Kamal HossainKamal Hossain · DataEngBytes Perth 2023
• 42mDeep Learning Recommender Architecture: Master Thesis to ...Igor Mazor · Berlin Buzzwords 2018
• 48mCollective IntelligencePaul Dix · GORUCO 2008
• 23mML-Powered Search & Recommendation 101Sergey Polyashov · Conf42 Machine Learning 2025
• 23mDeep Recommendations for e-CommerceGeorgios Damaskinos · Global Artificial Intelligence Virtual Conference 2022
• 27mMachine Learning for Advanced AI SearchSowjanya Pandruju · Conf42 Machine Learning 2025